Model validation plots
Eric R. Scott
24 August, 2021
TODO
- check that # of knots are adequate
- check diagnostics for individual-level random effects. Doesn’t look good, but not sure how to interpret or what to do.
Survival
Forest Fragments
Diagnostics
s_1ha_qres <- qresid(s_1ha)
par(mfrow = c(2,2))
#histogram
plot(density(s_1ha_qres))
#QQ plot
qqnorm(s_1ha_qres); qqline(s_1ha_qres)
#Fitted vs. residuals--binned residuals plot
binnedplot(fitted(s_1ha), residuals(s_1ha, type = "response"))
Basis dimensions
k.check(s_1ha)## k' edf k-index p-value
## s(log_size_prev) 9 2.842177e+00 0.9584417 0.935
## s(spei_history,L) 90 1.077927e+01 NA NA
## s(plot) 4 1.627924e-04 NA NABasis dimension checking with gam.check() doesn’t appear to work for dlnm crossbasis smooths. Instead I’ll use a method described in the help file ?choose.k to check for adequate knots. Unfortunately, bs = "cs" doesn’t work with dlnm, so I’ll use select = TRUE instead to reduce chance of overfitting.
check_res_edf(s_1ha)## # A tibble: 1 x 2
## smooth edf
## <chr> <dbl>
## 1 te(spei_history,L) 0The crossbasis smooth possibly needs more knots.
# plot_lag_slice(s_1ha, "spei_history", lag = 0) + labs(title = "surv, 1ha")Here you can see how the CIs hardly overlap 0, but it could be an artifact of the line not being allowed to be wiggly enough.
Summary
summary(s_1ha)##
## Family: binomial
## Link function: logit
##
## Formula:
## surv ~ flwr_prev + s(log_size_prev, bs = "cr", k = k[1]) + s(spei_history,
## L, bs = "cb", k = k[2:3], xt = list(bs = "cr")) + s(plot,
## bs = "re")
##
## Parametric coefficients:
## Estimate Std. Error z value Pr(>|z|)
## (Intercept) 3.22347 0.09128 35.312 <2e-16 ***
## flwr_prev1 -0.56668 0.44209 -1.282 0.2
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Approximate significance of smooth terms:
## edf Ref.df Chi.sq p-value
## s(log_size_prev) 2.842e+00 9 360.41 <2e-16 ***
## s(spei_history,L) 1.078e+01 22 73.09 <2e-16 ***
## s(plot) 1.628e-04 3 0.00 0.481
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## R-sq.(adj) = 0.0626 Deviance explained = 13.8%
## fREML = 12454 Scale est. = 1 n = 9183draw(s_1ha)## Warning: Removed 910 rows containing non-finite values (stat_contour).
Continuous Forest
Diagnostics
s_cf_qres <- qresid(s_cf)
par(mfrow = c(2,2))
#histogram
plot(density(s_cf_qres))
#QQ plot
qqnorm(s_cf_qres); qqline(s_cf_qres)
#Fitted vs. residuals--binned residuals plot
binnedplot(fitted(s_cf), residuals(s_cf, type = "response"))
Basis dimensions
k.check(s_cf)## k' edf k-index p-value
## s(log_size_prev) 9 3.457925 0.9099995 0.1925
## s(spei_history,L) 210 12.795976 NA NA
## s(plot) 6 4.336851 NA NA# looking for near zero edf
check_res_edf(s_cf)## # A tibble: 1 x 2
## smooth edf
## <chr> <dbl>
## 1 te(spei_history,L) 1.1Summary
summary(s_cf)##
## Family: binomial
## Link function: logit
##
## Formula:
## surv ~ flwr_prev + s(log_size_prev, bs = "cr", k = k[1]) + s(spei_history,
## L, bs = "cb", k = k[2:3], xt = list(bs = "cr")) + s(plot,
## bs = "re")
##
## Parametric coefficients:
## Estimate Std. Error z value Pr(>|z|)
## (Intercept) 3.4656 0.1277 27.147 <2e-16 ***
## flwr_prev1 0.1002 0.2685 0.373 0.709
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Approximate significance of smooth terms:
## edf Ref.df Chi.sq p-value
## s(log_size_prev) 3.458 9 1957.82 <2e-16 ***
## s(spei_history,L) 12.796 20 186.40 <2e-16 ***
## s(plot) 4.337 5 37.83 <2e-16 ***
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## R-sq.(adj) = 0.113 Deviance explained = 20%
## fREML = 43577 Scale est. = 1 n = 31701draw(s_cf)## Warning: Removed 939 rows containing non-finite values (stat_contour).
Effect of sample size
Here I use a random sub-sample to check that differences between continuous forest and fragments are not purely due to sample size differences, particularly differences in the complexity of the crossbasis smooth.
summary(s_1ha)$edf[3]## [1] 0.0001627924summary(s_cf)$edf[3]## [1] 4.336851summary(s_cf_sub)$edf[3]## [1] 3.346315draw(s_cf_sub)## Warning: Removed 910 rows containing non-finite values (stat_contour).
Surface still looks different in a similar way though.
Growth
Fragments
Diagnostics
appraise(g_1ha)
Basis dimensions
k.check(g_1ha)## k' edf k-index p-value
## s(log_size_prev) 9 4.193967 0.9818494 0.1025
## s(spei_history,L) 60 18.181554 NA NA
## s(plot) 4 2.826248 NA NAcheck_res_edf(g_1ha)## # A tibble: 1 x 2
## smooth edf
## <chr> <dbl>
## 1 te(spei_history,L) 0Summary
summary(g_1ha)##
## Family: Scaled t(4.729,0.477)
## Link function: identity
##
## Formula:
## log_size ~ flwr_prev + s(log_size_prev, bs = "cr", k = k[1]) +
## s(spei_history, L, bs = "cb", k = k[2:3], xt = list(bs = "cr")) +
## s(plot, bs = "re")
##
## Parametric coefficients:
## Estimate Std. Error t value Pr(>|t|)
## (Intercept) 4.17955 0.08340 50.117 <2e-16 ***
## flwr_prev1 0.08603 0.03601 2.389 0.0169 *
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Approximate significance of smooth terms:
## edf Ref.df F p-value
## s(log_size_prev) 4.194 9 3006.20 <2e-16 ***
## s(spei_history,L) 18.182 21 27.04 <2e-16 ***
## s(plot) 2.826 3 23.03 <2e-16 ***
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## R-sq.(adj) = 0.696 Deviance explained = 63.2%
## fREML = 12189 Scale est. = 1 n = 8527draw(g_1ha)## Warning: Removed 910 rows containing non-finite values (stat_contour).
Continuous Forest
Diagnostics
gratia::appraise(g_cf)
Basis dimensions
k.check(g_cf)## k' edf k-index p-value
## s(log_size_prev) 24 9.386988 0.967252 0.015
## s(spei_history,L) 60 15.144505 NA NA
## s(plot) 6 4.080451 NA NAcheck_res_edf(g_cf)## # A tibble: 1 x 2
## smooth edf
## <chr> <dbl>
## 1 te(spei_history,L) 0# plot_lag_slice(g_1ha, "spei_history", lag = 33)hmm… CIs are super narrow here, but effect size is tiny
Summary
summary(g_cf)##
## Family: Scaled t(3.852,0.414)
## Link function: identity
##
## Formula:
## log_size ~ flwr_prev + s(log_size_prev, bs = "cr", k = k[1]) +
## s(spei_history, L, bs = "cb", k = k[2:3], xt = list(bs = "cr")) +
## s(plot, bs = "re")
##
## Parametric coefficients:
## Estimate Std. Error t value Pr(>|t|)
## (Intercept) 4.36045 0.03401 128.214 <2e-16 ***
## flwr_prev1 0.02856 0.01476 1.934 0.0531 .
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Approximate significance of smooth terms:
## edf Ref.df F p-value
## s(log_size_prev) 9.387 24 6717.102 <2e-16 ***
## s(spei_history,L) 15.145 20 95.585 <2e-16 ***
## s(plot) 4.080 5 7.035 <2e-16 ***
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## R-sq.(adj) = 0.785 Deviance explained = 69.1%
## fREML = 41001 Scale est. = 1 n = 28849draw(g_cf)## Warning: Removed 968 rows containing non-finite values (stat_contour).
Continuous Forest subsample
To check that differences are not purely due to sample size differences, particularly that lower edf in continuous forests is due to higher sample size.
summary(g_1ha)$edf[3]## [1] 2.826248summary(g_cf)$edf[3]## [1] 4.080451summary(g_cf_sub)$edf[3]## [1] 4.82767edf of subsample is similar to full dataset, despite differences in sample size.
draw(g_cf_sub)## Warning: Removed 910 rows containing non-finite values (stat_contour).
Crossbasis surface looks nearly identical.
Flowering
Fragments
Diagnostics
f_1ha_qres <- qresid(f_1ha)
par(mfrow = c(2,2))
#histogram
plot(density(f_1ha_qres))
#QQ plot
qqnorm(f_1ha_qres); qqline(f_1ha_qres)
#Fitted vs. residuals--binned residuals plot
binnedplot(fitted(f_1ha), residuals(f_1ha, type = "response"))
Basis dimensions
k.check(f_1ha)## k' edf k-index p-value
## s(log_size_prev) 9 3.400302 0.998162 0.96
## s(spei_history,L) 162 14.127281 NA NA
## s(plot) 4 2.514545 NA NA# looking for near zero edf
check_res_edf(f_1ha)## # A tibble: 1 x 2
## smooth edf
## <chr> <dbl>
## 1 te(spei_history,L) 0Summary
summary(f_1ha)##
## Family: binomial
## Link function: logit
##
## Formula:
## flwr ~ flwr_prev + s(log_size_prev, bs = "cr", k = k[1]) + s(spei_history,
## L, bs = "cb", k = k[2:3], xt = list(bs = "cr")) + s(plot,
## bs = "re")
##
## Parametric coefficients:
## Estimate Std. Error z value Pr(>|z|)
## (Intercept) -6.2272 0.5207 -11.959 < 2e-16 ***
## flwr_prev1 1.0492 0.1685 6.225 4.8e-10 ***
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Approximate significance of smooth terms:
## edf Ref.df Chi.sq p-value
## s(log_size_prev) 3.400 9 321.86 < 2e-16 ***
## s(spei_history,L) 14.127 26 120.82 < 2e-16 ***
## s(plot) 2.515 3 16.44 0.000106 ***
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## R-sq.(adj) = 0.242 Deviance explained = 40%
## fREML = 11617 Scale est. = 1 n = 8527draw(f_1ha)## Warning: Removed 910 rows containing non-finite values (stat_contour).
Continuous Forest
Diagnostics
f_cf_qres <- qresid(f_cf)
par(mfrow = c(2,2))
#histogram
plot(density(f_cf_qres))
#QQ plot
qqnorm(f_cf_qres); qqline(f_cf_qres)
#Fitted vs. residuals--binned residuals plot
binnedplot(fitted(f_cf), residuals(f_cf, type = "response"))
Basis dimensions
k.check(f_cf)## k' edf k-index p-value
## s(log_size_prev) 9 5.795259 0.9746386 0.71
## s(spei_history,L) 210 11.451012 NA NA
## s(plot) 6 3.950101 NA NA# looking for near zero edf
check_res_edf(f_cf)## # A tibble: 1 x 2
## smooth edf
## <chr> <dbl>
## 1 te(spei_history,L) 1.2Summary
summary(f_cf)##
## Family: binomial
## Link function: logit
##
## Formula:
## flwr ~ flwr_prev + s(log_size_prev, bs = "cr", k = k[1]) + s(spei_history,
## L, bs = "cb", k = k[2:3], xt = list(bs = "cr")) + s(plot,
## bs = "re")
##
## Parametric coefficients:
## Estimate Std. Error z value Pr(>|z|)
## (Intercept) -5.07388 0.22648 -22.40 <2e-16 ***
## flwr_prev1 0.96103 0.08378 11.47 <2e-16 ***
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Approximate significance of smooth terms:
## edf Ref.df Chi.sq p-value
## s(log_size_prev) 5.795 9 1748.08 <2e-16 ***
## s(spei_history,L) 11.451 20 415.26 <2e-16 ***
## s(plot) 3.950 5 47.55 <2e-16 ***
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## R-sq.(adj) = 0.188 Deviance explained = 32.7%
## fREML = 41028 Scale est. = 1 n = 28849draw(f_cf)## Warning: Removed 968 rows containing non-finite values (stat_contour).
Continuous Forest subsample
To check that differences are not purely due to sample size differences, particularly that lower edf in continuous forests is due to higher sample size. For flowering, edf is actually slightly higher in CF with a larger sample size. With subsample, edf are more similar.
summary(f_1ha)$edf[3]## [1] 2.514545summary(f_cf)$edf[3]## [1] 3.950101summary(f_cf_sub)$edf[3]## [1] 4.860281draw(f_cf_sub)## Warning: Removed 910 rows containing non-finite values (stat_contour).
Crossbasis surface looks extremely similar.
Reproducibility
Reproducibility receipt
## datetime
Sys.time()## [1] "2021-08-24 11:46:23 EDT"## repository
if(requireNamespace('git2r', quietly = TRUE)) {
git2r::repository()
} else {
c(
system2("git", args = c("log", "--name-status", "-1"), stdout = TRUE),
system2("git", args = c("remote", "-v"), stdout = TRUE)
)
}## Local: revisions /Users/scottericr/Documents/HeliconiaDemography
## Remote: revisions @ origin (https://github.com/BrunaLab/HeliconiaDemography.git)
## Head: [b1aefac] 2021-08-24: move re smooth to end of formula, just for better printing of summary.## session info
sessionInfo()## R version 4.0.2 (2020-06-22)
## Platform: x86_64-apple-darwin17.0 (64-bit)
## Running under: macOS 10.16
##
## Matrix products: default
## BLAS: /Library/Frameworks/R.framework/Versions/4.0/Resources/lib/libRblas.dylib
## LAPACK: /Library/Frameworks/R.framework/Versions/4.0/Resources/lib/libRlapack.dylib
##
## locale:
## [1] en_US.UTF-8/en_US.UTF-8/en_US.UTF-8/C/en_US.UTF-8/en_US.UTF-8
##
## attached base packages:
## [1] stats4 parallel stats graphics grDevices datasets utils
## [8] methods base
##
## other attached packages:
## [1] arm_1.11-2 lme4_1.1-27.1 Matrix_1.3-3 MASS_7.3-54
## [5] qqplotr_0.0.5 Hmisc_4.5-0 Formula_1.2-4 survival_3.2-11
## [9] lattice_0.20-44 readxl_1.3.1 colorspace_2.0-1 rmarkdown_2.7
## [13] statmod_1.4.36 latex2exp_0.5.0 gratia_0.6.0.9112 broom_0.7.6
## [17] patchwork_1.1.1 glue_1.4.2 bbmle_1.0.23.1 dlnm_2.4.5
## [21] mgcv_1.8-36 nlme_3.1-152 lubridate_1.7.10 janitor_2.1.0
## [25] tsModel_0.6 SPEI_1.7 lmomco_2.3.6 tsibble_1.0.1
## [29] forcats_0.5.1 stringr_1.4.0 dplyr_1.0.5 purrr_0.3.4
## [33] readr_1.4.0 tidyr_1.1.3 tibble_3.1.1 ggplot2_3.3.5
## [37] tidyverse_1.3.1 here_1.0.1 tarchetypes_0.2.0 targets_0.4.2
## [41] dotenv_1.0.3 conflicted_1.0.4
##
## loaded via a namespace (and not attached):
## [1] backports_1.2.1 igraph_1.2.6 splines_4.0.2
## [4] digest_0.6.27 htmltools_0.5.1.1 fansi_0.4.2
## [7] magrittr_2.0.1 checkmate_2.0.0 memoise_2.0.0
## [10] cluster_2.1.2 modelr_0.1.8 bdsmatrix_1.3-4
## [13] anytime_0.3.9 jpeg_0.1-8.1 rvest_1.0.0
## [16] haven_2.4.1 xfun_0.22 callr_3.7.0
## [19] crayon_1.4.1 jsonlite_1.7.2 gtable_0.3.0
## [22] DEoptimR_1.0-8 abind_1.4-5 scales_1.1.1
## [25] mvtnorm_1.1-1 DBI_1.1.1 Rcpp_1.0.6
## [28] isoband_0.2.4 htmlTable_2.1.0 foreign_0.8-81
## [31] htmlwidgets_1.5.3 httr_1.4.2 RColorBrewer_1.1-2
## [34] ellipsis_0.3.2 pkgconfig_2.0.3 farver_2.1.0
## [37] nnet_7.3-16 sass_0.3.1 dbplyr_2.1.1
## [40] utf8_1.2.1 tidyselect_1.1.1 labeling_0.4.2
## [43] rlang_0.4.11 munsell_0.5.0 cellranger_1.1.0
## [46] tools_4.0.2 cachem_1.0.4 cli_2.5.0
## [49] generics_0.1.0 evaluate_0.14 fastmap_1.1.0
## [52] yaml_2.2.1 goftest_1.2-2 processx_3.5.2
## [55] knitr_1.33 fs_1.5.0 robustbase_0.93-7
## [58] mvnfast_0.2.5.1 xml2_1.3.2 compiler_4.0.2
## [61] rstudioapi_0.13 png_0.1-7 reprex_2.0.0
## [64] clustermq_0.8.95.1 bslib_0.2.4 stringi_1.6.2
## [67] highr_0.9 ps_1.6.0 nloptr_1.2.2.2
## [70] vctrs_0.3.8 pillar_1.6.0 lifecycle_1.0.0
## [73] jquerylib_0.1.4 data.table_1.14.0 R6_2.5.0
## [76] latticeExtra_0.6-29 renv_0.13.2 gridExtra_2.3
## [79] Lmoments_1.3-1 codetools_0.2-18 boot_1.3-28
## [82] assertthat_0.2.1 rprojroot_2.0.2 withr_2.4.2
## [85] hms_1.1.0 grid_4.0.2 rpart_4.1-15
## [88] coda_0.19-4 minqa_1.2.4 snakecase_0.11.0
## [91] git2r_0.28.0 numDeriv_2016.8-1.1 base64enc_0.1-3